Face Mask Detection With A Data Centric Approach For Training Deep Learning Model

Abstract

Millions have been killed and hundreds of millions have been infected by the SARS-CoV-2 virus that causes the COVID-19 pandemic. Facial masks prevent particles and droplets from spreading too far into the atmosphere, hence reducing transmission. Consequently, there is an increasing requirement for automated systems that can detect if individuals are not wearing masks or if they are wearing masks in the wrong way. This thesis topic was researched to address the issue referred to before by using the approach that focuses on data, called data-centric approach. Andrew Ng, who is generally acknowledged as the pioneer of modern artificial intelligence, endorsed this approach. According to him, a high-quality data collection and preparation process will have a critical influence on the quality of AI systems that are implemented in the future [1]. This research focuses on the development of the artificial intelligence systems utilizing the technique above, including the implementation of a pipeline that assures data quality prior to the model training process, which is critical in developing the models without focusing on the algorithms. As a result, the technique used in this research, which is data-centric, has the potential to increase the performance of the facial mask detection system dramatically, as about 85% in mAP and double in wAP respectively, compared to the baseline which is without data preprocessing.